Your recordings stay
on your machine.

Ibis watches a folder on your Windows PC and transcribes every audio file automatically, using a local Whisper model that runs entirely offline. Drop a voice memo. Get a text file. Nothing ever leaves your computer.

Cloud transcription services receive your recordings, process them on remote servers, and store them under terms of service you agreed to but probably did not read. The convenience is real. So is the trade. For people whose audio is sensitive by profession or by preference, the trade has never been acceptable. The local alternatives, until recently, were either bad or inaccessible.

Ibis is the alternative. It uses OpenAI’s Whisper large-v3-turbo model, packaged inside the installer. It runs on the CPU of an ordinary Windows machine. No GPU required. No account. No internet after install. The audio never moves.

Ibis lives in the system tray and watches a folder you choose. Every audio file you drop in becomes a transcript, automatically, while you do other things.

The setup is three steps. Point Ibis at a folder during installation. Drop in any MP3, WAV, or M4A file. A .txt transcript appears in a sibling Transcripts folder, timestamped and named to match the recording. A toast notification confirms the job is done. The original audio moves to Completed.

That is the entire interaction model. There is no application window to open, no batch to run, no queue to manage. The tool is invisible until it is useful.

The privacy claim is structural, not aspirational. The Whisper model lives inside the installer. The watch-folder process runs locally. The transcripts are written to your disk. Audio in the Completed folder is encrypted at rest using age v1 with X25519 keys. The secret key lives in the Windows Credential Manager, not in any file Ibis can read. We do not see it. We do not sync it.

The model bundle ships with all 99 languages Whisper supports. There is nothing to download per language. There is no opt-in pack. The whole multilingual capability is on disk after install, available offline forever.

On the standard public ASR benchmark, Ibis hits 1.94% word-error rate on LibriSpeech test-clean and 3.87% on test-other (harder audio with accents and noise), using Whisper large-v3-turbo with int8 quantization. Both numbers are within 0.27 percentage points of OpenAI’s published full-precision baselines.

A note on speed. Ibis v1.0 runs on CPU, so expect roughly real-time transcription on consumer hardware. A v1.0.1 release will detect available GPU and accelerate substantially. Accuracy is unaffected by the device choice; quantization (int8) is identical on CPU and CUDA.

Methodology, per-split breakdown, and full reproducibility commands are in the substantiation paper at doi.org/10.5281/zenodo.19931717.

Ibis was built for the people who keep encountering the same problem. Journalists protect sources and need transcripts of interviews that should not pass through a third-party server. Therapists and clinicians handle session recordings under HIPAA constraints that local-only processing makes much easier to satisfy. Attorneys work with privileged communications and depositions. Researchers bound by IRB protocols often cannot use cloud services for participant data. Knowledge workers want voice notes flowing into their second brains automatically. Podcasters and creators need transcripts for show notes and captions in seconds, offline.

The tool does not care which of these you are. The transcript appears, locally, with a timestamp and a filename. The audio stays where it was.

Ibis ships in three flavors. Personal is the only tier available now: $24, one-time, up to three devices, lifetime updates for v1.x. Pro is on the Q4 2026 roadmap at $49 with project workspaces and priority support; the Pro waitlist is open below. Team is a multi-seat bundle for practices and firms; reach out at ibis@simdadllc.com for terms.